By 2026 the consumer AI subscription had a magic number: $20 a month. ChatGPT Plus, Claude Pro, Perplexity Pro and Cursor Pro all landed there. It's a clean price — but on a flat $20/mo plan, every customer's token usage is a variable cost you carry. So what does $20 actually buy, and where does it stop being profitable?
This is a teardown of the price, not of anyone's books. We don't know any company's real model, usage, or negotiated rates — so we do the one thing public numbers allow: the break-even math. Given a public price and a model's public token rate, how much can a customer use before their cost equals what they pay?
What $20 buys in tokens
Take the output side alone — the priciest part of any LLM bill. At $20/mo, the break-even output volume is simply $20 ÷ the model's output rate:
- On Claude Sonnet ($15/Mtok output): about 1.3M output tokens — roughly a million words of generated text.
- On GPT-4o ($10/Mtok output): about 2.0M output tokens.
- On a reasoning model ($20/Mtok output): about 1.0M output tokens — and reasoning models also generate long hidden chains of thought you pay for.
- On GPT-4o mini ($0.60/Mtok output): about 33M output tokens — roughly 25× the headroom of Claude Sonnet for the same $20.
And that's the generous version: it ignores input tokens entirely. Add the prompts, context and retrieved documents a real product sends on every request, and every break-even number above comes down. A power user on a coding or agent product can burn through 1.3M tokens in days, not a month.
The model is the margin lever
That spread is the whole story: the same $20 plan is about 25× more forgiving on GPT-4o mini than on Claude Sonnet. This is why model routing — a cheap model for easy requests, the premium one only when it's needed — is the single biggest lever on whether a flat AI plan survives its heavy users. Pick the wrong default model and your most active customers become your biggest losses.
See it per model
The proof: GitHub blinked
If flat AI pricing always worked, the biggest players would keep it. They didn't. In June 2026, GitHub moved Copilot to usage-based billing — every plan now includes a monthly allotment of AI Credits, with paid usage beyond it. When the company behind one of the most efficient AI coding products on earth adds usage metering to a $10–39/mo plan, it's the clearest signal yet: at scale, flat pricing and heavy AI usage don't coexist without a margin mechanism.
What to do with this
You don't have to abandon flat pricing — but you do have to instrument it. Track gross margin per customer so you can see the heavy users before they hurt; know each plan's break-even usage; route easy traffic to a cheaper model; and add usage limits or higher tiers for the accounts that blow past break-even. Flat pricing is fine — flat pricing you can't see inside is not.
Run your own price point through the free calculator — plan price in, model and usage in, margin out — then connect Stripe and your LLM cost with MarginWard to watch every customer's margin automatically.
Methodology: prices are public list prices as of June 2026 (ChatGPT Plus, Claude Pro, Perplexity Pro and Cursor Pro at $20/mo; GitHub Copilot Pro $10, Business $19, Enterprise $39). Model rates are indicative public per-Mtok prices. Break-even = price ÷ output rate (output-only, an upper bound). No figure here is a claim about any company's real usage, model, or margin.